Research On Dual-Sensors Target Location Method For Complicated Scenes Applications

The accurate target location technology in complicated scenes is being applied to systems such as autonomous driving,which has important commercial and military value.Thesis is an academic and technical summary of related research work in this context.In simple scenarios with stable structured features,target location technology based on a single sensor can complete localization tasks.However,in complex scenes with sparse features,little semantic information or large location range,the data provided by a single sensor can not fully describe the complex features of the scene,which is likely to lead to location failure.Therefore,target positioning technology based on dual sensor data fusion has become a key research direction in practical process applications.Thesis studies a dual sensor target localization method for complicated scene applications using Li DAR and optical cameras.The common basic approach of the two methods is to use bimodal data obtained from Li DAR and cameras,then use data-driven methods to encode the sensing data,ultimately achieving similarity matching after feature encoding.The main work content and contributions of thesis are summarized as follows.Firstly,a target location method based on Li DAR is proposed.The main content is:first,Res Net based on Minkowski convolution is used to extract features;Then,a pyramid module incorporating adaptive features is used for fusion,which involves upsampling and downsampling of the feature map to achieve the fusion of features;Finally,a Ge M layer is used for pooling to generate descriptor can represent the target information scene.On this basis,the loss function of location similarity is designed to train the network.Secondly,propose a target location method based on dual-sensors.The main content is: first,design a descriptor extraction network,which uses Dense Net to extract features and generate feature maps,and uses Ge M layer to pool feature maps to obtain descriptors;Then designing a decision level fusion method to aggregate features of visual and point cloud descriptors,a bimodal descriptor is obtained;Finally,aiming at the overfitting phenomenon in training between different sensor data,the triple loss function is improved.Thirdly,experimental validation of our method was conducted on Oxford Robot Car dataset.Through comparative and ablation experiments,the values of average recall rate Top1 and average recall rate Top1% were compared.For the target location method based on Li DAR,experiments have shown that the average recall rates of network designed,Top1 and Top1%,have reached 94.2% and 98.1%,respectively,which is superior to point cloud target location methods such as Point Net VLAD.For dual-sensors target location method,experiments have shown that Top1 and Top1%,have reached 96.8% and 99.3%,respectively,indicating the effectiveness of the dual modal descriptor extraction network.